From 59de0c0bc70ea06c017333c518a8e9ae1f183f95 Mon Sep 17 00:00:00 2001
From: Debargha Mukherjee <debargha@google.com>
Date: Thu, 12 Nov 2015 17:14:32 -0800
Subject: [PATCH] Adding encoder support for 128x128 coding units

Changes to allow the encoder to make use of 128x128 coding units.

Change-Id: I340bd38f9d9750cb6346d83885efb00443852910
---
 test/vp9_subtract_test.cc           |   2 +-
 vp9/common/vp9_rtcd_defs.pl         | 216 +++++++++++++++++++++++++++-
 vp9/encoder/vp9_context_tree.c      |  13 ++
 vp9/encoder/vp9_encodeframe.c       |  94 ++++++++++--
 vp9/encoder/vp9_encoder.c           | 124 ++++++++++++++++
 vp9/encoder/vp9_rd.c                |   5 +-
 vp9/encoder/vp9_rdopt.c             |   2 +-
 vp9/encoder/vp9_sad.c               |  32 +++++
 vp9/encoder/vp9_variance.c          |  28 ++++
 vp9/encoder/x86/vp9_denoiser_sse2.c |   7 +
 10 files changed, 510 insertions(+), 13 deletions(-)

diff --git a/test/vp9_subtract_test.cc b/test/vp9_subtract_test.cc
index fabb43824..92dc476b5 100644
--- a/test/vp9_subtract_test.cc
+++ b/test/vp9_subtract_test.cc
@@ -91,7 +91,7 @@ TEST_P(VP9SubtractBlockTest, SimpleSubtract) {
 INSTANTIATE_TEST_CASE_P(C, VP9SubtractBlockTest,
                         ::testing::Values(vp9_subtract_block_c));
 
-#if HAVE_SSE2 && CONFIG_USE_X86INC
+#if HAVE_SSE2 && CONFIG_USE_X86INC && !CONFIG_EXT_CODING_UNIT_SIZE
 INSTANTIATE_TEST_CASE_P(SSE2, VP9SubtractBlockTest,
                         ::testing::Values(vp9_subtract_block_sse2));
 #endif
diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 27dfb36bb..e7e4e9a80 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -1058,6 +1058,17 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
 
 
 # variance
+if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+add_proto qw/unsigned int vp9_variance128x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+specialize qw/vp9_variance128x128/;
+
+add_proto qw/unsigned int vp9_variance128x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+specialize qw/vp9_variance128x64/;
+
+add_proto qw/unsigned int vp9_variance64x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+specialize qw/vp9_variance64x128/;
+}
+
 add_proto qw/unsigned int vp9_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_variance32x16 avx2/, "$sse2_x86inc";
 
@@ -1103,6 +1114,26 @@ specialize qw/vp9_variance4x8/, "$sse2_x86inc";
 add_proto qw/unsigned int vp9_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_variance4x4/, "$sse2_x86inc";
 
+if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+add_proto qw/unsigned int vp9_sub_pixel_variance128x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+specialize qw/vp9_sub_pixel_variance128x128/;
+
+add_proto qw/unsigned int vp9_sub_pixel_avg_variance128x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+specialize qw/vp9_sub_pixel_avg_variance128x128/;
+
+add_proto qw/unsigned int vp9_sub_pixel_variance128x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+specialize qw/vp9_sub_pixel_variance128x64/;
+
+add_proto qw/unsigned int vp9_sub_pixel_avg_variance128x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+specialize qw/vp9_sub_pixel_avg_variance128x64/;
+
+add_proto qw/unsigned int vp9_sub_pixel_variance64x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+specialize qw/vp9_sub_pixel_variance64x128/;
+
+add_proto qw/unsigned int vp9_sub_pixel_avg_variance64x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+specialize qw/vp9_sub_pixel_avg_variance64x128/;
+}
+
 add_proto qw/unsigned int vp9_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_sub_pixel_variance64x64 avx2/, "$sse2_x86inc", "$ssse3_x86inc";
 
@@ -1183,6 +1214,17 @@ specialize qw/vp9_sub_pixel_variance4x4/, "$sse_x86inc", "$ssse3_x86inc";
 add_proto qw/unsigned int vp9_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
 specialize qw/vp9_sub_pixel_avg_variance4x4/, "$sse_x86inc", "$ssse3_x86inc";
 
+if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+add_proto qw/unsigned int vp9_sad128x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
+specialize qw/vp9_sad128x128/;
+
+add_proto qw/unsigned int vp9_sad128x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vp9_sad128x64/;
+
+add_proto qw/unsigned int vp9_sad64x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vp9_sad64x128/;
+}
+
 add_proto qw/unsigned int vp9_sad64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
 specialize qw/vp9_sad64x64 neon avx2/, "$sse2_x86inc";
 
@@ -1222,6 +1264,17 @@ specialize qw/vp9_sad4x8/, "$sse_x86inc";
 add_proto qw/unsigned int vp9_sad4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
 specialize qw/vp9_sad4x4/, "$sse_x86inc";
 
+if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+add_proto qw/unsigned int vp9_sad128x128_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
+specialize qw/vp9_sad128x128_avg/;
+
+add_proto qw/unsigned int vp9_sad128x64_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vp9_sad128x64_avg/;
+
+add_proto qw/unsigned int vp9_sad64x128_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+specialize qw/vp9_sad64x128_avg/;
+}
+
 add_proto qw/unsigned int vp9_sad64x64_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad64x64_avg avx2/, "$sse2_x86inc";
 
@@ -1261,6 +1314,11 @@ specialize qw/vp9_sad4x8_avg/, "$sse_x86inc";
 add_proto qw/unsigned int vp9_sad4x4_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
 specialize qw/vp9_sad4x4_avg/, "$sse_x86inc";
 
+if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+add_proto qw/void vp9_sad128x128x3/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array";
+specialize qw/vp9_sad128x128x3/;
+}
+
 add_proto qw/void vp9_sad64x64x3/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array";
 specialize qw/vp9_sad64x64x3/;
 
@@ -1282,6 +1340,11 @@ specialize qw/vp9_sad8x8x3 sse3/;
 add_proto qw/void vp9_sad4x4x3/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array";
 specialize qw/vp9_sad4x4x3 sse3/;
 
+if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+add_proto qw/void vp9_sad128x128x8/, "const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array";
+specialize qw/vp9_sad128x128x8/;
+}
+
 add_proto qw/void vp9_sad64x64x8/, "const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array";
 specialize qw/vp9_sad64x64x8/;
 
@@ -1309,6 +1372,17 @@ specialize qw/vp9_sad4x8x8/;
 add_proto qw/void vp9_sad4x4x8/, "const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array";
 specialize qw/vp9_sad4x4x8 sse4/;
 
+if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+add_proto qw/void vp9_sad128x128x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
+specialize qw/vp9_sad128x128x4d/;
+
+add_proto qw/void vp9_sad128x64x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
+specialize qw/vp9_sad128x64x4d/;
+
+add_proto qw/void vp9_sad64x128x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
+specialize qw/vp9_sad64x128x4d/;
+}
+
 add_proto qw/void vp9_sad64x64x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
 specialize qw/vp9_sad64x64x4d sse2 avx2/;
 
@@ -1492,9 +1566,14 @@ if (vpx_config("CONFIG_WEDGE_PARTITION") eq "yes") {
 }
 # ENCODEMB INVOKE
 
+if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
 add_proto qw/void vp9_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
-specialize qw/vp9_subtract_block neon/, "$sse2_x86inc";
+specialize qw/vp9_subtract_block/;
 
+}else{
+add_proto qw/void vp9_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
+specialize qw/vp9_subtract_block neon/, "$sse2_x86inc";
+}
 #
 # Denoiser
 #
@@ -1745,6 +1824,16 @@ specialize qw/vp9_temporal_filter_apply sse2/;
 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 
   # variance
+  if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+  add_proto qw/unsigned int vp9_highbd_variance128x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_variance128x128/;
+
+  add_proto qw/unsigned int vp9_highbd_variance128x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_variance128x64/;
+
+  add_proto qw/unsigned int vp9_highbd_variance64x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_variance64x128/;
+  }
   add_proto qw/unsigned int vp9_highbd_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
   specialize qw/vp9_highbd_variance32x16/, "$sse2_x86inc";
 
@@ -1790,6 +1879,17 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vp9_highbd_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
   specialize qw/vp9_highbd_get16x16var/, "$sse2_x86inc";
 
+  if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+  add_proto qw/unsigned int vp9_highbd_10_variance128x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_10_variance128x128/;
+
+  add_proto qw/unsigned int vp9_highbd_10_variance128x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_10_variance128x64/;
+
+  add_proto qw/unsigned int vp9_highbd_10_variance64x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_10_variance64x128/;
+  }
+
   add_proto qw/unsigned int vp9_highbd_10_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
   specialize qw/vp9_highbd_10_variance32x16/, "$sse2_x86inc";
 
@@ -1835,6 +1935,17 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vp9_highbd_10_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
   specialize qw/vp9_highbd_10_get16x16var/, "$sse2_x86inc";
 
+  if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+  add_proto qw/unsigned int vp9_highbd_12_variance128x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_12_variance128x128/;
+
+  add_proto qw/unsigned int vp9_highbd_12_variance128x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_12_variance128x64/;
+
+  add_proto qw/unsigned int vp9_highbd_12_variance64x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_12_variance64x128/;
+  }
+
   add_proto qw/unsigned int vp9_highbd_12_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
   specialize qw/vp9_highbd_12_variance32x16/, "$sse2_x86inc";
 
@@ -1880,6 +1991,26 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vp9_highbd_12_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
   specialize qw/vp9_highbd_12_get16x16var/, "$sse2_x86inc";
 
+  if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+  add_proto qw/unsigned int vp9_highbd_sub_pixel_variance128x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_sub_pixel_variance128x128/;
+
+  add_proto qw/unsigned int vp9_highbd_sub_pixel_avg_variance128x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_sub_pixel_avg_variance128x128/;
+
+  add_proto qw/unsigned int vp9_highbd_sub_pixel_variance128x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_sub_pixel_variance128x64/;
+
+  add_proto qw/unsigned int vp9_highbd_sub_pixel_avg_variance128x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_sub_pixel_avg_variance128x64/;
+
+  add_proto qw/unsigned int vp9_highbd_sub_pixel_variance64x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_sub_pixel_variance64x128/;
+
+  add_proto qw/unsigned int vp9_highbd_sub_pixel_avg_variance64x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_sub_pixel_avg_variance64x128/;
+  }
+
   add_proto qw/unsigned int vp9_highbd_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
   specialize qw/vp9_highbd_sub_pixel_variance64x64/, "$sse2_x86inc";
 
@@ -1958,6 +2089,26 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/unsigned int vp9_highbd_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
   specialize qw/vp9_highbd_sub_pixel_avg_variance4x4/;
 
+  if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+  add_proto qw/unsigned int vp9_highbd_10_sub_pixel_variance128x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_10_sub_pixel_variance128x128/;
+
+  add_proto qw/unsigned int vp9_highbd_10_sub_pixel_avg_variance128x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_10_sub_pixel_avg_variance128x128/;
+
+  add_proto qw/unsigned int vp9_highbd_10_sub_pixel_variance128x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_10_sub_pixel_variance128x64/;
+
+  add_proto qw/unsigned int vp9_highbd_10_sub_pixel_avg_variance128x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_10_sub_pixel_avg_variance128x64/;
+
+  add_proto qw/unsigned int vp9_highbd_10_sub_pixel_variance64x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_10_sub_pixel_variance64x128/;
+
+  add_proto qw/unsigned int vp9_highbd_10_sub_pixel_avg_variance64x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_10_sub_pixel_avg_variance64x128/;
+  }
+
   add_proto qw/unsigned int vp9_highbd_10_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
   specialize qw/vp9_highbd_10_sub_pixel_variance64x64/, "$sse2_x86inc";
 
@@ -2036,6 +2187,26 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/unsigned int vp9_highbd_10_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
   specialize qw/vp9_highbd_10_sub_pixel_avg_variance4x4/;
 
+  if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+  add_proto qw/unsigned int vp9_highbd_12_sub_pixel_variance128x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_12_sub_pixel_variance128x128/;
+
+  add_proto qw/unsigned int vp9_highbd_12_sub_pixel_avg_variance128x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_12_sub_pixel_avg_variance128x128/;
+
+  add_proto qw/unsigned int vp9_highbd_12_sub_pixel_variance128x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_12_sub_pixel_variance128x64/;
+
+  add_proto qw/unsigned int vp9_highbd_12_sub_pixel_avg_variance128x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_12_sub_pixel_avg_variance128x64/;
+
+  add_proto qw/unsigned int vp9_highbd_12_sub_pixel_variance64x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/vp9_highbd_12_sub_pixel_variance64x128/;
+
+  add_proto qw/unsigned int vp9_highbd_12_sub_pixel_avg_variance64x128/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_12_sub_pixel_avg_variance64x128/;
+  }
+
   add_proto qw/unsigned int vp9_highbd_12_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
   specialize qw/vp9_highbd_12_sub_pixel_variance64x64/, "$sse2_x86inc";
 
@@ -2114,6 +2285,17 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/unsigned int vp9_highbd_12_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
   specialize qw/vp9_highbd_12_sub_pixel_avg_variance4x4/;
 
+  if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+  add_proto qw/unsigned int vp9_highbd_sad128x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
+  specialize qw/vp9_highbd_sad128x128/;
+
+  add_proto qw/unsigned int vp9_highbd_sad128x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vp9_highbd_sad128x64/;
+
+  add_proto qw/unsigned int vp9_highbd_sad64x128/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
+  specialize qw/vp9_highbd_sad64x128/;
+  }
+
   add_proto qw/unsigned int vp9_highbd_sad64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
   specialize qw/vp9_highbd_sad64x64/, "$sse2_x86inc";
 
@@ -2153,6 +2335,17 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/unsigned int vp9_highbd_sad4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
   specialize qw/vp9_highbd_sad4x4/;
 
+  if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+  add_proto qw/unsigned int vp9_highbd_sad128x128_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_sad128x128_avg/;
+
+  add_proto qw/unsigned int vp9_highbd_sad128x64_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_sad128x64_avg/;
+
+  add_proto qw/unsigned int vp9_highbd_sad64x128_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+  specialize qw/vp9_highbd_sad64x128_avg/;
+  }
+
   add_proto qw/unsigned int vp9_highbd_sad64x64_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
   specialize qw/vp9_highbd_sad64x64_avg/, "$sse2_x86inc";
 
@@ -2192,6 +2385,11 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/unsigned int vp9_highbd_sad4x4_avg/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred";
   specialize qw/vp9_highbd_sad4x4_avg/;
 
+  if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+  add_proto qw/void vp9_highbd_sad128x128x3/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array";
+  specialize qw/vp9_highbd_sad128x128x3/;
+  }
+
   add_proto qw/void vp9_highbd_sad64x64x3/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array";
   specialize qw/vp9_highbd_sad64x64x3/;
 
@@ -2213,6 +2411,11 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vp9_highbd_sad4x4x3/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array";
   specialize qw/vp9_highbd_sad4x4x3/;
 
+  if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+  add_proto qw/void vp9_highbd_sad128x128x8/, "const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array";
+  specialize qw/vp9_highbd_sad128x128x8/;
+  }
+
   add_proto qw/void vp9_highbd_sad64x64x8/, "const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array";
   specialize qw/vp9_highbd_sad64x64x8/;
 
@@ -2240,6 +2443,17 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vp9_highbd_sad4x4x8/, "const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array";
   specialize qw/vp9_highbd_sad4x4x8/;
 
+  if (vpx_config("CONFIG_EXT_CODING_UNIT_SIZE") eq "yes") {
+  add_proto qw/void vp9_highbd_sad128x128x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
+  specialize qw/vp9_highbd_sad128x128x4d/;
+
+  add_proto qw/void vp9_highbd_sad128x64x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
+  specialize qw/vp9_highbd_sad128x64x4d/;
+
+  add_proto qw/void vp9_highbd_sad64x128x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
+  specialize qw/vp9_highbd_sad64x128x4d/;
+  }
+
   add_proto qw/void vp9_highbd_sad64x64x4d/, "const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array";
   specialize qw/vp9_highbd_sad64x64x4d sse2/;
 
diff --git a/vp9/encoder/vp9_context_tree.c b/vp9/encoder/vp9_context_tree.c
index de8c2c2dd..ea76d63f1 100644
--- a/vp9/encoder/vp9_context_tree.c
+++ b/vp9/encoder/vp9_context_tree.c
@@ -16,6 +16,9 @@ static const BLOCK_SIZE square[] = {
   BLOCK_16X16,
   BLOCK_32X32,
   BLOCK_64X64,
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  BLOCK_128X128,
+#endif
 };
 
 static void alloc_mode_context(VP9_COMMON *cm, int num_4x4_blk,
@@ -144,8 +147,13 @@ static void free_tree_contexts(PC_TREE *tree) {
 // represents the state of our search.
 void vp9_setup_pc_tree(VP9_COMMON *cm, VP9_COMP *cpi) {
   int i, j;
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  const int leaf_nodes = 64 * 4;
+  const int tree_nodes = 64 * 4 + 64 + 16 + 4 + 1;
+#else
   const int leaf_nodes = 64;
   const int tree_nodes = 64 + 16 + 4 + 1;
+#endif
   int pc_tree_index = 0;
   PC_TREE *this_pc;
   PICK_MODE_CONTEXT *this_leaf;
@@ -199,8 +207,13 @@ void vp9_setup_pc_tree(VP9_COMMON *cm, VP9_COMP *cpi) {
 }
 
 void vp9_free_pc_tree(VP9_COMP *cpi) {
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  const int tree_nodes = 64 * 4 + 64 + 16 + 4 + 1;
+  const int leaf_nodes = 64 * 4;
+#else
   const int tree_nodes = 64 + 16 + 4 + 1;
   const int leaf_nodes = 64;
+#endif
   int i;
 
   // Set up all 4x4 mode contexts
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index 989099219..37cd81feb 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -114,6 +114,16 @@ static const uint8_t VP9_VAR_OFFS[CODING_UNIT_SIZE] = {
     128, 128, 128, 128, 128, 128, 128, 128,
     128, 128, 128, 128, 128, 128, 128, 128,
     128, 128, 128, 128, 128, 128, 128, 128,
+#if CONFIG_EXT_CODING_UNIT_SIZE
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+#endif
 };
 
 #if CONFIG_VP9_HIGHBITDEPTH
@@ -126,6 +136,16 @@ static const uint16_t VP9_HIGH_VAR_OFFS_8[CODING_UNIT_SIZE] = {
     128, 128, 128, 128, 128, 128, 128, 128,
     128, 128, 128, 128, 128, 128, 128, 128,
     128, 128, 128, 128, 128, 128, 128, 128,
+#if CONFIG_EXT_CODING_UNIT_SIZE
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128,
+    128, 128, 128, 128, 128, 128, 128, 128
+#endif
 };
 
 static const uint16_t VP9_HIGH_VAR_OFFS_10[CODING_UNIT_SIZE] = {
@@ -137,6 +157,16 @@ static const uint16_t VP9_HIGH_VAR_OFFS_10[CODING_UNIT_SIZE] = {
     128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
     128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
     128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+#if CONFIG_EXT_CODING_UNIT_SIZE
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
+    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4
+#endif
 };
 
 static const uint16_t VP9_HIGH_VAR_OFFS_12[CODING_UNIT_SIZE] = {
@@ -148,6 +178,16 @@ static const uint16_t VP9_HIGH_VAR_OFFS_12[CODING_UNIT_SIZE] = {
     128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
     128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
     128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+#if CONFIG_EXT_CODING_UNIT_SIZE
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
+    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16
+#endif
 };
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 
@@ -432,6 +472,13 @@ typedef struct {
   v32x32 split[4];
 } v64x64;
 
+#if CONFIG_EXT_CODING_UNIT_SIZE
+typedef struct {
+  partition_variance part_variances;
+  v64x64 split[4];
+} v128x128;
+#endif
+
 typedef struct {
   partition_variance *part_variances;
   var *split[4];
@@ -441,6 +488,9 @@ typedef enum {
   V16X16,
   V32X32,
   V64X64,
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  V128X128,
+#endif
 } TREE_LEVEL;
 
 static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
@@ -448,6 +498,15 @@ static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
   node->part_variances = NULL;
   vpx_memset(node->split, 0, sizeof(node->split));
   switch (bsize) {
+#if CONFIG_EXT_CODING_UNIT_SIZE
+    case BLOCK_128X128: {
+      v128x128 *vt = (v128x128 *) data;
+      node->part_variances = &vt->part_variances;
+      for (i = 0; i < 4; i++)
+        node->split[i] = &vt->split[i].part_variances.none;
+      break;
+    }
+#endif
     case BLOCK_64X64: {
       v64x64 *vt = (v64x64 *) data;
       node->part_variances = &vt->part_variances;
@@ -600,6 +659,10 @@ static void choose_partitioning(VP9_COMP *cpi,
   vp9_clear_system_state();
   vp9_zero(vt);
   set_offsets(cpi, tile, mi_row, mi_col, BLOCK_LARGEST);
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  printf("Not yet implemented: choose_partitioning\n");
+  exit(-1);
+#endif
 
   if (xd->mb_to_right_edge < 0)
     pixels_wide += (xd->mb_to_right_edge >> 3);
@@ -2024,8 +2087,8 @@ static void rd_use_partition(VP9_COMP *cpi, const TileInfo *const tile,
   int i, pl;
   PARTITION_TYPE partition = PARTITION_NONE;
   BLOCK_SIZE subsize;
-  ENTROPY_CONTEXT l[(CODING_UNIT_SIZE / 4) * MAX_MB_PLANE];
-  ENTROPY_CONTEXT a[(CODING_UNIT_SIZE / 4) * MAX_MB_PLANE];
+  ENTROPY_CONTEXT l[(CODING_UNIT_SIZE >> 2) * MAX_MB_PLANE];
+  ENTROPY_CONTEXT a[(CODING_UNIT_SIZE >> 2) * MAX_MB_PLANE];
   PARTITION_CONTEXT sl[CODING_UNIT_SIZE >> 3], sa[CODING_UNIT_SIZE >> 3];
   RD_COST last_part_rdc, none_rdc, chosen_rdc;
   BLOCK_SIZE sub_subsize = BLOCK_4X4;
@@ -2275,8 +2338,8 @@ static void rd_use_partition(VP9_COMP *cpi, const TileInfo *const tile,
 #if CONFIG_SUPERTX
       int rt_nocoef = 0;
 #endif
-      ENTROPY_CONTEXT l[(CODING_UNIT_SIZE / 4) * MAX_MB_PLANE];
-      ENTROPY_CONTEXT a[(CODING_UNIT_SIZE / 4) * MAX_MB_PLANE];
+      ENTROPY_CONTEXT l[(CODING_UNIT_SIZE >> 2) * MAX_MB_PLANE];
+      ENTROPY_CONTEXT a[(CODING_UNIT_SIZE >> 2) * MAX_MB_PLANE];
       PARTITION_CONTEXT sl[CODING_UNIT_SIZE >> 3], sa[CODING_UNIT_SIZE >> 3];
 
       if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
@@ -2385,7 +2448,10 @@ static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
   BLOCK_4X4,   BLOCK_4X4,   BLOCK_4X4,
   BLOCK_8X8,   BLOCK_8X8,   BLOCK_8X8,
   BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
-  BLOCK_16X16
+  BLOCK_16X16,
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  BLOCK_16X16, BLOCK_16X16, BLOCK_16X16
+#endif
 };
 
 static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
@@ -2393,7 +2459,10 @@ static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
   BLOCK_16X16, BLOCK_32X32, BLOCK_32X32,
   BLOCK_32X32, BLOCK_64X64, BLOCK_64X64,
   BLOCK_64X64, BLOCK_64X64, BLOCK_64X64,
-  BLOCK_64X64
+  BLOCK_64X64,
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  BLOCK_64X64, BLOCK_64X64, BLOCK_128X128
+#endif
 };
 
 // Look at all the mode_info entries for blocks that are part of this
@@ -2431,7 +2500,10 @@ static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
   BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
   BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
   BLOCK_32X32, BLOCK_32X32, BLOCK_32X32,
-  BLOCK_64X64
+  BLOCK_64X64,
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  BLOCK_64X64, BLOCK_64X64, BLOCK_128X128
+#endif
 };
 
 // Look at neighboring blocks and set a min and max partition size based on
@@ -2869,8 +2941,8 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
   const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2;
-  ENTROPY_CONTEXT l[(CODING_UNIT_SIZE / 4) * MAX_MB_PLANE];
-  ENTROPY_CONTEXT a[(CODING_UNIT_SIZE / 4) * MAX_MB_PLANE];
+  ENTROPY_CONTEXT l[(CODING_UNIT_SIZE >> 2) * MAX_MB_PLANE];
+  ENTROPY_CONTEXT a[(CODING_UNIT_SIZE >> 2) * MAX_MB_PLANE];
   PARTITION_CONTEXT sl[CODING_UNIT_SIZE >> 3], sa[CODING_UNIT_SIZE >> 3];
   TOKENEXTRA *tp_orig = *tp;
   PICK_MODE_CONTEXT *ctx = &pc_tree->none;
@@ -3799,7 +3871,11 @@ static void encode_rd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
   MACROBLOCKD *const xd = &cpi->mb.e_mbd;
   SPEED_FEATURES *const sf = &cpi->sf;
   int mi_col;
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  const int leaf_nodes = 64 * 4;
+#else
   const int leaf_nodes = 64;
+#endif
 
   // Initialize the left context for the new SB row
   vpx_memset(&xd->left_context, 0, sizeof(xd->left_context));
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 0de65bfcb..138e21f76 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -779,6 +779,19 @@ static void fnname##_bits12(const uint8_t *src_ptr, \
   sad_array[i] >>= 4; \
 }
 
+#if CONFIG_EXT_CODING_UNIT_SIZE
+MAKE_BFP_SAD_WRAPPER(vp9_highbd_sad128x128)
+MAKE_BFP_SADAVG_WRAPPER(vp9_highbd_sad128x128_avg)
+MAKE_BFP_SAD3_WRAPPER(vp9_highbd_sad128x128x3)
+MAKE_BFP_SAD8_WRAPPER(vp9_highbd_sad128x128x8)
+MAKE_BFP_SAD4D_WRAPPER(vp9_highbd_sad128x128x4d)
+MAKE_BFP_SAD_WRAPPER(vp9_highbd_sad128x64)
+MAKE_BFP_SADAVG_WRAPPER(vp9_highbd_sad128x64_avg)
+MAKE_BFP_SAD4D_WRAPPER(vp9_highbd_sad128x64x4d)
+MAKE_BFP_SAD_WRAPPER(vp9_highbd_sad64x128)
+MAKE_BFP_SADAVG_WRAPPER(vp9_highbd_sad64x128_avg)
+MAKE_BFP_SAD4D_WRAPPER(vp9_highbd_sad64x128x4d)
+#endif
 MAKE_BFP_SAD_WRAPPER(vp9_highbd_sad32x16)
 MAKE_BFP_SADAVG_WRAPPER(vp9_highbd_sad32x16_avg)
 MAKE_BFP_SAD4D_WRAPPER(vp9_highbd_sad32x16x4d)
@@ -1020,6 +1033,38 @@ static void  highbd_set_var_fns(VP9_COMP *const cpi) {
                    vp9_highbd_sad4x4x8_bits8,
                    vp9_highbd_sad4x4x4d_bits8)
 
+#if CONFIG_EXT_CODING_UNIT_SIZE
+        HIGHBD_BFP(BLOCK_128X128,
+                   vp9_highbd_sad128x128_bits8,
+                   vp9_highbd_sad128x128_avg_bits8,
+                   vp9_highbd_variance128x128,
+                   vp9_highbd_sub_pixel_variance128x128,
+                   vp9_highbd_sub_pixel_avg_variance128x128,
+                   vp9_highbd_sad128x128x3_bits8,
+                   vp9_highbd_sad128x128x8_bits8,
+                   vp9_highbd_sad128x128x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_128X64,
+                   vp9_highbd_sad128x64_bits8,
+                   vp9_highbd_sad128x64_avg_bits8,
+                   vp9_highbd_variance128x64,
+                   vp9_highbd_sub_pixel_variance128x64,
+                   vp9_highbd_sub_pixel_avg_variance128x64,
+                   NULL,
+                   NULL,
+                   vp9_highbd_sad128x64x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_64X128,
+                   vp9_highbd_sad64x128_bits8,
+                   vp9_highbd_sad64x128_avg_bits8,
+                   vp9_highbd_variance64x128,
+                   vp9_highbd_sub_pixel_variance64x128,
+                   vp9_highbd_sub_pixel_avg_variance64x128,
+                   NULL,
+                   NULL,
+                   vp9_highbd_sad64x128x4d_bits8)
+#endif
+
 #if CONFIG_WEDGE_PARTITION
         HIGHBD_MBFP(BLOCK_64X64,
                     vp9_highbd_masked_sad64x64_bits8,
@@ -1207,6 +1252,38 @@ static void  highbd_set_var_fns(VP9_COMP *const cpi) {
                    vp9_highbd_sad4x4x8_bits10,
                    vp9_highbd_sad4x4x4d_bits10)
 
+#if CONFIG_EXT_CODING_UNIT_SIZE
+        HIGHBD_BFP(BLOCK_128X128,
+                   vp9_highbd_sad128x128_bits10,
+                   vp9_highbd_sad128x128_avg_bits10,
+                   vp9_highbd_10_variance128x128,
+                   vp9_highbd_10_sub_pixel_variance128x128,
+                   vp9_highbd_10_sub_pixel_avg_variance128x128,
+                   vp9_highbd_sad128x128x3_bits10,
+                   vp9_highbd_sad128x128x8_bits10,
+                   vp9_highbd_sad128x128x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_128X64,
+                   vp9_highbd_sad128x64_bits10,
+                   vp9_highbd_sad128x64_avg_bits10,
+                   vp9_highbd_10_variance128x64,
+                   vp9_highbd_10_sub_pixel_variance128x64,
+                   vp9_highbd_10_sub_pixel_avg_variance128x64,
+                   NULL,
+                   NULL,
+                   vp9_highbd_sad128x64x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_64X128,
+                   vp9_highbd_sad64x128_bits10,
+                   vp9_highbd_sad64x128_avg_bits10,
+                   vp9_highbd_10_variance64x128,
+                   vp9_highbd_10_sub_pixel_variance64x128,
+                   vp9_highbd_10_sub_pixel_avg_variance64x128,
+                   NULL,
+                   NULL,
+                   vp9_highbd_sad64x128x4d_bits10)
+#endif
+
 #if CONFIG_WEDGE_PARTITION
         HIGHBD_MBFP(BLOCK_64X64,
                     vp9_highbd_masked_sad64x64_bits10,
@@ -1394,6 +1471,38 @@ static void  highbd_set_var_fns(VP9_COMP *const cpi) {
                    vp9_highbd_sad4x4x8_bits12,
                    vp9_highbd_sad4x4x4d_bits12)
 
+#if CONFIG_EXT_CODING_UNIT_SIZE
+        HIGHBD_BFP(BLOCK_128X128,
+                   vp9_highbd_sad128x128_bits12,
+                   vp9_highbd_sad128x128_avg_bits12,
+                   vp9_highbd_12_variance128x128,
+                   vp9_highbd_12_sub_pixel_variance128x128,
+                   vp9_highbd_12_sub_pixel_avg_variance128x128,
+                   vp9_highbd_sad128x128x3_bits12,
+                   vp9_highbd_sad128x128x8_bits12,
+                   vp9_highbd_sad128x128x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_128X64,
+                   vp9_highbd_sad128x64_bits12,
+                   vp9_highbd_sad128x64_avg_bits12,
+                   vp9_highbd_12_variance128x64,
+                   vp9_highbd_12_sub_pixel_variance128x64,
+                   vp9_highbd_12_sub_pixel_avg_variance128x64,
+                   NULL,
+                   NULL,
+                   vp9_highbd_sad128x64x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_64X128,
+                   vp9_highbd_sad64x128_bits12,
+                   vp9_highbd_sad64x128_avg_bits12,
+                   vp9_highbd_12_variance64x128,
+                   vp9_highbd_12_sub_pixel_variance64x128,
+                   vp9_highbd_12_sub_pixel_avg_variance64x128,
+                   NULL,
+                   NULL,
+                   vp9_highbd_sad64x128x4d_bits12)
+#endif
+
 #if CONFIG_WEDGE_PARTITION
         HIGHBD_MBFP(BLOCK_64X64,
                     vp9_highbd_masked_sad64x64_bits12,
@@ -1815,6 +1924,21 @@ VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf) {
     cpi->fn_ptr[BT].sdx8f          = SDX8F; \
     cpi->fn_ptr[BT].sdx4df         = SDX4DF;
 
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  BFP(BLOCK_128X128, vp9_sad128x128, vp9_sad128x128_avg,
+      vp9_variance128x128, vp9_sub_pixel_variance128x128,
+      vp9_sub_pixel_avg_variance128x128, vp9_sad128x128x3, vp9_sad128x128x8,
+      vp9_sad128x128x4d)
+
+  BFP(BLOCK_128X64, vp9_sad128x64, vp9_sad128x64_avg,
+      vp9_variance128x64, vp9_sub_pixel_variance128x64,
+      vp9_sub_pixel_avg_variance128x64, NULL, NULL, vp9_sad128x64x4d)
+
+  BFP(BLOCK_64X128, vp9_sad64x128, vp9_sad64x128_avg,
+      vp9_variance64x128, vp9_sub_pixel_variance64x128,
+      vp9_sub_pixel_avg_variance64x128, NULL, NULL, vp9_sad64x128x4d)
+#endif
+
   BFP(BLOCK_32X16, vp9_sad32x16, vp9_sad32x16_avg,
       vp9_variance32x16, vp9_sub_pixel_variance32x16,
       vp9_sub_pixel_avg_variance32x16, NULL, NULL, vp9_sad32x16x4d)
diff --git a/vp9/encoder/vp9_rd.c b/vp9/encoder/vp9_rd.c
index 2179ba42d..e81032ce7 100644
--- a/vp9/encoder/vp9_rd.c
+++ b/vp9/encoder/vp9_rd.c
@@ -61,7 +61,10 @@ void vp9_rd_cost_init(RD_COST *rd_cost) {
 // This table is used to correct for block size.
 // The factors here are << 2 (2 = x0.5, 32 = x8 etc).
 static const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES] = {
-  2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32
+  2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32,
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  48, 48, 64
+#endif
 };
 
 static void fill_mode_costs(VP9_COMP *cpi) {
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index 16835ce51..0597c2cb0 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -7910,7 +7910,7 @@ void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
     TX_SIZE best_tx_size;
     int rate2_tx, this_skip2_tx = 0;
     int64_t distortion2_tx, bestrd_tx = INT64_MAX;
-    uint8_t tmp_zcoeff_blk[(CODING_UNIT_SIZE * CODING_UNIT_SIZE) / 16];
+    uint8_t tmp_zcoeff_blk[(CODING_UNIT_SIZE * CODING_UNIT_SIZE) >> 4];
 #endif  // CONFIG_EXT_TX
 
     *mbmi = *inter_ref_list[copy_mode - REF0];
diff --git a/vp9/encoder/vp9_sad.c b/vp9/encoder/vp9_sad.c
index f926a5500..9081dd765 100644
--- a/vp9/encoder/vp9_sad.c
+++ b/vp9/encoder/vp9_sad.c
@@ -66,6 +66,22 @@ void vp9_sad##m##x##n##x4d_c(const uint8_t *src, int src_stride, \
     sads[i] = vp9_sad##m##x##n##_c(src, src_stride, refs[i], ref_stride); \
 }
 
+#if CONFIG_EXT_CODING_UNIT_SIZE
+// 128x128
+sadMxN(128, 128)
+sadMxNxK(128, 128, 3)
+sadMxNxK(128, 128, 8)
+sadMxNx4D(128, 128)
+
+// 128x64
+sadMxN(128, 64)
+sadMxNx4D(128, 64)
+
+// 64x128
+sadMxN(64, 128)
+sadMxNx4D(64, 128)
+#endif
+
 // 64x64
 sadMxN(64, 64)
 sadMxNxK(64, 64, 3)
@@ -205,6 +221,22 @@ void vp9_highbd_sad##m##x##n##x4d_c(const uint8_t *src, int src_stride, \
   }  \
 }
 
+#if CONFIG_EXT_CODING_UNIT_SIZE
+// 128x128
+highbd_sadMxN(128, 128)
+highbd_sadMxNxK(128, 128, 3)
+highbd_sadMxNxK(128, 128, 8)
+highbd_sadMxNx4D(128, 128)
+
+// 128x64
+highbd_sadMxN(128, 64)
+highbd_sadMxNx4D(128, 64)
+
+// 64x128
+highbd_sadMxN(64, 128)
+highbd_sadMxNx4D(64, 128)
+#endif
+
 // 64x64
 highbd_sadMxN(64, 64)
 highbd_sadMxNxK(64, 64, 3)
diff --git a/vp9/encoder/vp9_variance.c b/vp9/encoder/vp9_variance.c
index 36f02da3f..ad0cc36ee 100644
--- a/vp9/encoder/vp9_variance.c
+++ b/vp9/encoder/vp9_variance.c
@@ -253,6 +253,20 @@ VAR(64, 64)
 SUBPIX_VAR(64, 64)
 SUBPIX_AVG_VAR(64, 64)
 
+#if CONFIG_EXT_CODING_UNIT_SIZE
+VAR(64, 128)
+SUBPIX_VAR(64, 128)
+SUBPIX_AVG_VAR(64, 128)
+
+VAR(128, 64)
+SUBPIX_VAR(128, 64)
+SUBPIX_AVG_VAR(128, 64)
+
+VAR(128, 128)
+SUBPIX_VAR(128, 128)
+SUBPIX_AVG_VAR(128, 128)
+#endif
+
 void vp9_comp_avg_pred(uint8_t *comp_pred, const uint8_t *pred, int width,
                        int height, const uint8_t *ref, int ref_stride) {
   int i, j;
@@ -632,6 +646,20 @@ HIGHBD_VAR(64, 64)
 HIGHBD_SUBPIX_VAR(64, 64)
 HIGHBD_SUBPIX_AVG_VAR(64, 64)
 
+#if CONFIG_EXT_CODING_UNIT_SIZE
+HIGHBD_VAR(64, 128)
+HIGHBD_SUBPIX_VAR(64, 128)
+HIGHBD_SUBPIX_AVG_VAR(64, 128)
+
+HIGHBD_VAR(128, 64)
+HIGHBD_SUBPIX_VAR(128, 64)
+HIGHBD_SUBPIX_AVG_VAR(128, 64)
+
+HIGHBD_VAR(128, 128)
+HIGHBD_SUBPIX_VAR(128, 128)
+HIGHBD_SUBPIX_AVG_VAR(128, 128)
+#endif
+
 void vp9_highbd_comp_avg_pred(uint16_t *comp_pred, const uint8_t *pred8,
                               int width, int height, const uint8_t *ref8,
                               int ref_stride) {
diff --git a/vp9/encoder/x86/vp9_denoiser_sse2.c b/vp9/encoder/x86/vp9_denoiser_sse2.c
index bf5fa889f..1a83f9ccd 100644
--- a/vp9/encoder/x86/vp9_denoiser_sse2.c
+++ b/vp9/encoder/x86/vp9_denoiser_sse2.c
@@ -460,9 +460,16 @@ int vp9_denoiser_filter_sse2(const uint8_t *sig, int sig_stride,
                                  avg, avg_stride,
                                  increase_denoising,
                                  bs, motion_magnitude);
+#if CONFIG_EXT_CODING_UNIT_SIZE
+  } else if (bs == BLOCK_16X8 || bs == BLOCK_16X16 || bs == BLOCK_16X32 ||
+             bs == BLOCK_32X16 || bs == BLOCK_32X32 || bs == BLOCK_32X64 ||
+             bs == BLOCK_64X32 || bs == BLOCK_64X64 || bs == BLOCK_64X128 ||
+             bs == BLOCK_128X64 || bs == BLOCK_128X128) {
+#else
   } else if (bs == BLOCK_16X8 || bs == BLOCK_16X16 || bs == BLOCK_16X32 ||
              bs == BLOCK_32X16|| bs == BLOCK_32X32 || bs == BLOCK_32X64 ||
              bs == BLOCK_64X32 || bs == BLOCK_64X64) {
+#endif
     return vp9_denoiser_64_32_16xM_sse2(sig, sig_stride,
                                         mc_avg, mc_avg_stride,
                                         avg, avg_stride,
-- 
2.40.0